Transforming Electrocardiogram to Participant-Level Diagnosis in Atrial Fibrillation Screening*

In recent developments, single-lead ECGs can now be captured outside clinical environments, opening opportunities for atrial fibrillation (AF) detection in the general population. However, this introduces significant challenges, including a very low AF prevalence rate in the captured datasets and a large number of ECGs per participant, which complicates the task of diagnosing participants effectively.In this study, a dynamic training approach which trained models to detect AF from ECGs, was introduced to incorporate more non-AF ECGs during training compared to a fixed approach, and the performance of both approaches was compared in terms of ECG classification and participant diagnosis. The two approaches were compared by separately training an RR-interval-based AF detection model using the Screening for Atrial Fibrillation with ECG to Reduce Stroke (SAFER) dataset. The participant-level testing was derived from two straightforward transitions by applying the ECG classification model.Results showed that the dynamic strategy consistently outperformed the fixed strategy, achieving higher AUROC (0.92-0.94 versus 0.90-0.92) and AUPRC scores (0.38-0.75 versus 0.12-0.66) on both ECG-level and participant-level testing tasks.Clinical relevance- Models in this study were trained using the Screening for Atrial Fibrillation with ECG to Reduce Stroke (SAFER) dataset, which was real-world data captured during the screening stage of a clinical trial. The findings suggest that the dynamic training approach not only advances the accuracy of ECG-level classification, but also demonstrates its applicability at the participant-level diagnosis, which is a crucial step toward real-world clinical applications.